PaperlayerView exact revision on arXiv

Source-linked AI summary

Long-Term Mobile Traffic Forecasting Using Deep Spatio-Temporal Neural Networks

Chaoyun Zhang, Paul Patras

arXiv:1712.08083v1cs.NI

TL;DR

The paper addresses accurate long-term, network-wide mobile traffic forecasting despite costly measurements and strong spatio-temporal variability. It proposes STN, OTS, and D-STN to learn spatio-temporal features, operate with limited ground truth, and combine predictions with historical statistics. On 60-day urban and rural datasets, the methods produced accurate 10-hour forecasts, reducing errors by up to 61% while using measurement intervals up to 600 times shorter.

  • Problem

    Accurate long-term network-wide mobile traffic forecasting is difficult because dense-deployment measurements are complex and expensive, while traffic varies spatio-temporally with user mobility.

  • Method

    The paper combines a ConvLSTM/3D-ConvNet STN with Ouroboros fine-tuning and a Double STN that blends predictions with historical statistics.

  • Results

    61% lower prediction errors and 600 times shorter ground-truth measurement durations were achieved against widely employed ARIMA and HW-ExpS methods.

  • Takeaways & Limitations

    The proposed (D-)STN schemes provide accurate 10-hour mobile traffic forecasts from short observation intervals across urban and rural datasets.

  • Takeaways & Limitations

    Uncertainty may grow over time, limiting accuracy for long-term forecasts exceeding 5 hours without the further D-STN treatment.

Abstract

from arXiv · show

Forecasting with high accuracy the volume of data traffic that mobile users will consume is becoming increasingly important for precision traffic engineering, demand-aware network resource allocation, as well as public transportation. Measurements collection in dense urban deployments is however complex and expensive, and the post-processing required to make predictions is highly non-trivial, given the intricate spatio-temporal variability of mobile traffic due to user mobility. To overcome these challenges, in this paper we harness the exceptional feature extraction abilities of deep learning and propose a Spatio-Temporal neural Network (STN) architecture purposely designed for precise network-wide mobile traffic forecasting. We present a mechanism that fine tunes the STN and enables its operation with only limited ground truth observations. We then introduce a Double STN technique (D-STN), which uniquely combines the STN predictions with historical statistics, thereby making faithful long-term mobile traffic projections. Experiments we conduct with real-world mobile traffic data sets, collected over 60 days in both urban and rural areas, demonstrate that the proposed (D-)STN schemes perform up to 10-hour long predictions with remarkable accuracy, irrespective of the time of day when they are triggered. Specifically, our solutions achieve up to 61% smaller prediction errors as compared to widely used forecasting approaches, while operating with up to 600 times shorter measurement intervals.

1 INTRODUCTION

Mobile traffic forecasting is important but difficult because monitoring is costly and traffic varies across locations and time. The paper proposes deep spatio-temporal models for long-term, network-wide forecasts from limited observations.

  • Motivation: Mobile traffic growth makes precision traffic engineering and demand-aware cellular resource allocation increasingly important.Future applications are expected to contribute substantially to demand.
  • Limitations of prior methods: Existing mobile prediction methods mainly model individual base-station time series and overlook spatial correlations caused by user movement.This limits their suitability for network-wide and longer-term forecasting.
  • Proposed approach: The STN fuses ConvLSTM and 3D-ConvNet features to model long-term trends, short-term variations, and traffic correlations across locations and times.An MLP maps the fused representation to future traffic volumes.
  • Proposed approach: The Ouroboros Training Scheme fine-tunes the network so earlier predictions can replace unavailable ground-truth observations during long-term forecasting.Training combines actual measurements with one-step predictions and minimizes their output difference from ground truth.
  • Proposed approach: The Double STN combines STN predictions with a decay mechanism and empirical local traffic means to mitigate accumulating errors and extend forecast duration.The combination keeps forecasts within reasonable bounds.
  • Results: 61% lower NRMSE and 600 times shorter measurement intervals were achieved against commonly used and traditional forecasting methods.Experiments used real-world data collected over 60 days and produced 10-hour forecasts from 2-hour observations.

2 THE MOBILE TRAFFIC FORECASTING PROBLEM

The forecasting problem represents network-wide mobile traffic as a spatio-temporal sequence over a geographical grid and predicts future snapshots from prior observations. The proposed STN uses an encoder-decoder architecture to learn spatial and temporal features from traffic matrices.

  • Problem formulation: Network-wide mobile traffic is represented as snapshots over time on an X × Y geographical grid.Each snapshot records traffic volume in square cells, forming a tensor D ∈ R^T×X×Y.
  • Problem formulation: The task is to predict the most likely K-step future sequence given S previous observations.The formulation conditions future traffic snapshots on the preceding observed sequence.
  • Spatial structure: Traffic values at a target cell depend largely on neighbouring cells, while distant-cell information can be neglected.The paper therefore confines the input region to adjacent cells around the target.
  • STN architecture: The STN follows an encoder-decoder design that processes traffic matrices with ConvLSTMs and 3D-ConvNets before decoding predictions through an MLP.The architecture is designed to learn spatio-temporal features from the input sequence.

3 THE SPATIO-TEMPORAL NETWORK

The STN combines ConvLSTM and 3D-ConvNet components in an encoder-decoder architecture to forecast mobile traffic from spatio-temporal traffic matrices. Its fusion and training design captures long-term trends, local fluctuations, and spatial correlations for multi-step forecasting.

  • Architecture: The STN uses an encoder-decoder architecture that processes traffic matrices through ConvLSTM and 3D-ConvNet elements before an MLP produces predictions.The encoder fuses learned feature maps, while the decoder uses fully connected layers for regression.
  • Architecture: ConvLSTMs capture long-term traffic trends and spatial correlations through convolutional hidden maps and gated spatio-temporal updates.Convolution replaces dense recurrent connections, reducing parameters while preserving spatial structure.
  • Architecture: 3D-ConvNets capture cross-temporal dependencies and local fluctuations without back-propagation through time.Their temporal convolutions improve modeling of short-term traffic variations and generalisation.
  • STN – Fusing ConvLSTMs and 3D-ConvNets: The STN fuses ConvLSTM and 3D-ConvNet outputs twice, combining long-term trends with local fluctuations for more robust predictions.The fusion operation averages intermediate outputs from both models before decoding.
  • Prediction: The MLP uses outputs across every time step, allowing predictions to incorporate all learned temporal features rather than only the final state.The STN shares weights between inputs, requiring fewer parameters than traditional RBM-based forecasting approaches.
  • Prediction: Each input contains 12 snapshots spanning 2 hours over a 10×10 spatial neighborhood, and the STN predicts the centre cell at the next time step.The resulting input tensor is 11 × 11 × 12, with 10-minute measurement intervals.

4 LONG-TERM MOBILE TRAFFIC FORECASTING

Long-term forecasting becomes difficult when ground-truth measurements disappear because recursively reused predictions accumulate errors. The paper addresses this with Ouroboros retraining and D-STN blending of predictions with historical traffic statistics.

  • Forecasting challenge: Ground-truth unavailability causes recursively reused prediction errors to accumulate during multi-step forecasting.The pure STN follows traffic accurately with measurements but rapidly fails to track actual evolution when measurements are suspended.
  • An Ouroboros Training Scheme: OTS retrains the STN with recursively generated predictions so the model learns to operate when earlier outputs replace unavailable ground truth.The training queue begins with S ground-truth observations, then repeatedly removes the oldest frame and appends the latest prediction before SGD training.
  • An Ouroboros Training Scheme: OTS broadens the model’s input support and suppresses the pure STN’s overestimation tendency by bringing retrained predictions closer to ground truth.The paper characterizes OTS as a prediction-based data augmentation technique.
  • Blending Predictions and Historical Statistics: OTS improves multi-step accuracy, but growing uncertainty can limit accuracy for long forecasts exceeding 5 hours.This motivates the further improved D-STN forecasting system.
  • Blending Predictions and Historical Statistics: D-STN combines STN predictions with weekly empirical means through a decay mechanism that reduces prediction weight over the forecast horizon.The sigmoid weighting function makes empirical means increasingly dominant as the horizon grows; the empirical mean can also be updated online.
  • Blending Predictions and Historical Statistics: D-STN normalizes its blended output while dynamically weighting the original and OTS-trained STN predictions.The original STN receives heavier weight initially because its inputs still partly contain ground-truth observations.

5 PERFORMANCE EVALUATION

The evaluation compares STN and D-STN with conventional and deep-learning predictors on real-world Milan and Trentino traffic data. Across long horizons, the proposed methods deliver accurate network-wide forecasts from short observations and generalize across locations.

  • Experimental setup: Experiments use Telecom Italia mobile traffic datasets from Milan and Trentino, evaluating multi-step forecasts against HW-ExpS, ARIMA, MLP, ConvLSTM, 3D-ConvNet, SVM, and AE+LSTM.Deep-learning models are trained on Milan data, while HW-ExpS and ARIMA are trained separately on both datasets.
  • Experimental setup: 2 hours of observations support 10-hour predictions for the deep-learning methods, whereas HW-ExpS and ARIMA continue receiving ground-truth data during forecasting.The two-hour observations are prediction inputs rather than training data.
  • Milan results: D-STN performs best across Milan prediction instances, with NRMSE up to 60% lower than HW-ExpS, 38% lower than ARIMA, and 26% lower than AE+LSTM for one-step prediction.It also outperforms individually used ConvLSTM, 3D-ConvNet, and MLP models.
  • Trentino and long-term results: 61% and 35% lower NRMSE than HW-ExpS and ARIMA, respectively, are achieved by STN in long-term predictions, while D-STN remains nearly as accurate in Trentino.ARIMA is slightly more accurate in short-term Milan predictions, where traffic is low, but STN performs substantially better over longer horizons.
  • Network-wide predictions: D-STN provides the best 10-hour network-wide Milan snapshot, while STN performs well in the city centre but slightly underestimates surrounding traffic.HW-ExpS overestimates traffic, whereas ARIMA, MLP, ConvLSTM, and 3D-ConvNet capture spatial profiles but misestimate volume.

6 RELATED WORK

Prior mobile traffic forecasting largely uses linear time-series methods or models individual base-station series, while deep-learning work supplies tools for sequential and spatio-temporal feature learning. This paper applies ConvLSTM and 3D-ConvNet capabilities to long-term mobile traffic prediction with limited observations.

  • Mobile traffic forecasting: Exponential smoothing and ARIMA are widely used linear time-series approaches for understanding and predicting mobile traffic dynamics.Earlier work includes Holt-Winters exponential smoothing for short-term GSM/GPRS forecasting.
  • Mobile traffic forecasting: Exploratory Factor Analysis has been used to analyze non-trivial spatio-temporal mobile traffic patterns for network activity profiling and land-use detection.
  • Deep-learning predictors: LSTM, ConvLSTM, and 3D-ConvNet architectures provide established approaches for sequential and spatio-temporal feature learning in tasks such as time-series prediction, precipitation nowcasting, and video recognition.
  • Deep-learning predictors: This work exploits ConvLSTM and 3D-ConvNet to predict mobile traffic for up to 10 hours using limited two-hour observations.

7 CONCLUSIONS

The paper presents STN-based cellular traffic forecasting to address the growing demand for timely resource allocation despite increasingly costly monitoring. OTS fine-tunes the model for limited observations, while D-STN combines STN outputs with historical statistics and achieves lower errors with much shorter measurement durations.

  • 7 CONCLUSIONS: The STN is designed to forecast cellular traffic precisely while leveraging deep neural networks to overcome prior forecasting limitations.The architecture targets network traffic forecasting for dynamic resource allocation and timely responses to growing demand.
  • 7 CONCLUSIONS: 61% lower prediction errors are achieved by the proposed (D-)STN schemes than by ARIMA and HW-ExpS methods.The comparison uses publicly available 60-day measurements from Milan and the Trentino region.
  • 7 CONCLUSIONS: 600 times shorter ground truth measurement durations are required by the proposed (D-)STN schemes.This supports forecasting when accurate monitoring is costly in densifying cellular deployments.
  • 7 CONCLUSIONS: OTS fine-tunes the pretrained model, and D-STN combines STN outputs with historical statistics to improve long-term prediction performance.These mechanisms address forecasting with limited ground truth observations and support the paper’s long-term forecasting objective.
Loading 1712.08083v1…